From 8f5f2dd5a70f47322614eda6f97304808447199c Mon Sep 17 00:00:00 2001
From: notaz <notasas@gmail.com>
Date: Sat, 10 Jan 2015 03:51:44 +0200
Subject: [PATCH] some random improvements

---
 plugins/dfsound/spu.c | 14 ++++++++------
 plugins/gpulib/gpu.c  | 12 ++++++++++++
 2 files changed, 20 insertions(+), 6 deletions(-)

diff --git a/plugins/dfsound/spu.c b/plugins/dfsound/spu.c
index f5edd3a7..ec31b0ca 100644
--- a/plugins/dfsound/spu.c
+++ b/plugins/dfsound/spu.c
@@ -595,24 +595,26 @@ make_do_samples(simple, , ,
 static int do_samples_skip(int ch, int ns_to)
 {
  SPUCHAN *s_chan = &spu.s_chan[ch];
+ int spos = s_chan->spos;
+ int sinc = s_chan->sinc;
  int ret = ns_to, ns, d;
 
- s_chan->spos += s_chan->iSBPos << 16;
+ spos += s_chan->iSBPos << 16;
 
  for (ns = 0; ns < ns_to; ns++)
  {
-  s_chan->spos += s_chan->sinc;
-  while (s_chan->spos >= 28*0x10000)
+  spos += sinc;
+  while (spos >= 28*0x10000)
   {
    d = skip_block(ch);
    if (d && ns < ret)
     ret = ns;
-   s_chan->spos -= 28*0x10000;
+   spos -= 28*0x10000;
   }
  }
 
- s_chan->iSBPos = s_chan->spos >> 16;
- s_chan->spos &= 0xffff;
+ s_chan->iSBPos = spos >> 16;
+ s_chan->spos = spos & 0xffff;
 
  return ret;
 }
diff --git a/plugins/gpulib/gpu.c b/plugins/gpulib/gpu.c
index 337e27a7..c9b05d4d 100644
--- a/plugins/gpulib/gpu.c
+++ b/plugins/gpulib/gpu.c
@@ -13,8 +13,16 @@
 #include "gpu.h"
 
 #define ARRAY_SIZE(x) (sizeof(x) / sizeof((x)[0]))
+#ifdef __GNUC__
 #define unlikely(x) __builtin_expect((x), 0)
+#define preload __builtin_prefetch
 #define noinline __attribute__((noinline))
+#else
+#define unlikely(x)
+#define preload(...)
+#define noinline
+#error huh
+#endif
 
 #define gpu_log(fmt, ...) \
   printf("%d:%03d: " fmt, *gpu.state.frame_count, *gpu.state.hcnt, ##__VA_ARGS__)
@@ -518,6 +526,8 @@ long GPUdmaChain(uint32_t *rambase, uint32_t start_addr)
   int len, left, count;
   long cpu_cycles = 0;
 
+  preload(rambase + (start_addr & 0x1fffff) / 4);
+
   if (unlikely(gpu.cmd_len > 0))
     flush_cmd_buffer();
 
@@ -537,6 +547,8 @@ long GPUdmaChain(uint32_t *rambase, uint32_t start_addr)
     list = rambase + (addr & 0x1fffff) / 4;
     len = list[0] >> 24;
     addr = list[0] & 0xffffff;
+    preload(rambase + (addr & 0x1fffff) / 4);
+
     cpu_cycles += 10;
     if (len > 0)
       cpu_cycles += 5 + len;
-- 
2.39.2